FPGA Programming: HLS4ML
by
Abstract
This course explores the hls4ml package for translating python models to hls code for deployment on FPGAs.
Sections of this course include:
- Basics for introduction to the hls4ml package and simple configurations such as model precision and their impact of firmware resources
- hl4ml options for exploring concepts such as reuse factor or strategy and their impact on resources
- Architectures for introduction on how different architectures can impact resource usage and some specific hls4ml configuration options for this
- Quantisation for an introduction to quantised aware training and how this is used in conjuction with hls4ml
- Inputs for a demonstration of the impact of changing the number of inputs in your ML model and how these scale with your model's resource usage
- Advanced architecturs for exploring more complex architecture types such as MLPmixer, GNNs and Transformers and how these can be tranlsated into FPGA firmware
- Decision trees for using Conifer to translage decision tree ML models into FPGA firmware
Lecturer Bio
I am a research fellow working on the CMS Level-1 trigger and design machine learning algorithms to make sub microsecond decisions whether to keep or discard incoming data. I specifically work on jet tagging algorithms and explore a range of different architectures and techniques to boost tagging performance while also reducing their inference latency and overall cost in hardware.
I am also interested in the impact of changing environments on trigger algorithms and how we can design robust machine learning or use continual learning to constantly update algorithms as the detector changes. I explore how different architectural considerations and learning techniques can change how an ML algorithm understands its training data and how that can help reduce the impact of noise in incoming data.
A further research area is in tree methods for fast machine learning, designing and implementing decision trees in firmware as well as method for updating them on the fly or self-updating learning.
Tutor Bio - Dr. Sioni Paris Summers (CERN)
I’m an applied physicst at CERN working on the Level 1 Trigger of the CMS Experiment at the LHC. For the Phase 2 Upgrade I develop fast and efficient physics reconstruction algorithms for FPGAs. I’ve contributed to track finding, vertex reconstruction, particle flow, pileup subtraction (PUPPI), jet reconstruction, electron identification, and providing a platform for particle-based algorithms.
I also develop algorithms and tools for machine learning in the trigger. I helped to deploy the first ML seed in the CMS trigger during LHC Run 3 in 2024, the anomaly detection trigger AXOL1TL. For the Phase 2 Upgrade of CMS I’m involved in several projects to enhance our data taking with ML, like jet tagging. I’m a task leader in the Next Generation Triggers project at CERN, leading efforts on deeply integrating ML into the CMS Phase 2 Trigger. On the ML tools side I created and maintain conifer for BDT inference on FPGAs, and I’ve previously been coordinator of hls4ml for NN inference on FPGAs.
I’m motivated to apply the techniques and technology we develop at CERN for fast ML inference in other areas as well. At the moment I’m working on a project, called Edge SpAIce, to use these tools for the detection of plastics pollution in the ocean onboard Earth Observation satellites. I’ve previously worked on a collaboration with autonomous vehicle company Zenseact.